The reduction described in the mentioned post, from SAT to the halting problem, performs the following mapping:
$$\varphi\mapsto \big(\langle M\rangle,\varphi \big)$$
where $\varphi$ is a CNF formula and $M$ is the Turing machine which given a CNF $\psi$ as input, iterates over all the possible assignments to the variables of $\psi$, and enters a loop iff no satisfying assignment was found.
The machine $M$ does not depend on the input of the reduction, so it has some constant size description independent of the input $\varphi$. This means that you can output $\langle M \rangle$ in polynomial (even constant) time relative to the size of $\varphi$, hence the reduction is computable in polynomial time.
Note that you only need to show that you can generate $\langle M \rangle$ in polynomial time, which does not necessarily mean it has to have a constant size. Suppose e.g. that we want to show that $L=\{\langle M\rangle | \text{M halts on input $\epsilon$}\}$ is NP hard. In that case, the reduction $\varphi\mapsto \langle M_{\varphi}\rangle$ does the trick, where $M_{\varphi}$ is the machine which ignores its input, iterates over all possible assignments for $\varphi$, and enters a loop if no satisfying assignment was found. $M_{\varphi}$ does not have a constant size description, since you have to hardcode $\varphi$ into it. However, you can still generate $\langle M_{\varphi}\rangle$ in polynomial time, which makes the reduction polynomial.
